OpenAI's latest reasoning models are hallucinating more than ever

New AI models are supposed to get smarter and lie less. OpenAI’s latest releases just broke that unspoken rule.

Despite hitting state-of-the-art marks in coding and math, OpenAI’s new reasoning models, o3 and o4-mini, hallucinate significantly more than their predecessors. Even more concerning, OpenAI admitted in its technical report that it doesn’t quite know why.

The smarter they think, the more they make up

OpenAI’s internal benchmarks paint a rough picture. On PersonQA, an internal test measuring facts about people, o3 hallucinated on 33% of questions.

That is roughly double the rate of o1 at 16% and o3-mini at 14.8%. The smaller o4-mini fared even worse, hallucinating on 48% of responses—essentially guessing half the time.

Both models hallucinate more often than older reasoning models and standard models like GPT-4o. OpenAI’s technical report notes that because these models make more claims overall, they output more accurate statements alongside more inaccurate ones. The company simply stated that more research is needed.

Outside labs are catching it in real time, too. Nonprofit research lab Transluce caught o3 hallucinating its own physical process. In one test, the model claimed it ran code outside ChatGPT on a 2021 MacBook Pro and pasted the numbers back in. The model has no such capability.

Neil Chowdhury, a researcher at Transluce and former OpenAI employee, suspects the reinforcement learning used in the o-series might amplify issues usually kept in check by standard post-training pipelines.

Meanwhile, Stanford adjunct professor and Workera CEO Kian Katanforoosh noted that while o3 beats rivals in coding tasks, it constantly hallucinates broken web links.

Why it matters

We think this highlights a critical blind spot in the industry’s pivot to reasoning models. Over the past year, labs rushed to test-time compute and reinforcement learning after traditional pre-training hit diminishing returns. But if scaling reasoning makes models act like overconfident fabulists, that trade-off is brutal.

For enterprise users, this makes models like o3 a tough sell outside sandbox environments. In software engineering, a compiler catches hallucinations instantly. But in legal, finance, or customer service, a model that makes up facts one-third of the time is a non-starter.

OpenAI points out that GPT-4o with web search hits 90% accuracy on SimpleQA, suggesting search retrieval could patch the problem. Still, if the underlying reasoning mechanism actively manufactures convincing lies, relying on search feels like a band-aid over a core architectural defect.